TextRNN 的 tensorflow-1.14 实现

RNN 原理自己找，这里只给出简单例子的实现代码来展示原理

import tensorflow as tf
import numpy as np

tf.reset_default_graph()
sentences = [‘i love damao’,’i like mengjun’,’we love all’]
words = list(set(” “.join(sentences).split()))
word2idx = {v:k for k,v in enumerate(words)}
idx2word = {k:v for k,v in enumerate(words)}

V = len(words) # 词典大小
step = 2 # 时间序列长度
hidden = 5 # 隐层大小
dim = 50 # 词向量维度

# 制作输入和标签
def make_batch(sentences):
input_batch = []
target_batch = []
for sentence in sentences:
words = sentence.split()
input = [word2idx[word] for word in words[:-1]]
target = word2idx[words[-1]]

input_batch.append(input)
target_batch.append(np.eye(V)[target]) # 这里将标签改为 one-hot 编码，之后计算交叉熵的时候会用到
return input_batch, target_batch

# 初始化词向量
embedding = tf.get_variable(shape=[V, dim], initializer=tf.random_normal_initializer(), name=”embedding”)

X = tf.placeholder(tf.int32, [None, step])
XX = tf.nn.embedding_lookup(embedding, X)
Y = tf.placeholder(tf.int32, [None, V])

# 定义 cell
cell = tf.nn.rnn_cell.BasicRNNCell(hidden)
# 计算各个时间点的输出和隐层输出的结果
outputs, hiddens = tf.nn.dynamic_rnn(cell, XX, dtype=tf.float32) # outputs: [batch_size, step, hidden] hiddens: [batch_size, hidden]

# 这里将所有时间点的状态向量都作为了后续分类器的输入（也可以只将*后时间节点的状态向量作为后续分类器的输入）
W = tf.Variable(tf.random_normal([step*hidden, V]))
b = tf.Variable(tf.random_normal([V]))
L = tf.matmul(tf.reshape(outputs,[-1, step*hidden]), W) + b

# 计算损失并进行优化
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y, logits=L))
optimizer = tf.train.AdamOptimizer(0.001).minimize(cost)

# 预测
prediction = tf.argmax(L, 1)

# 初始化 tf
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)

# 喂训练数据
input_batch, target_batch = make_batch(sentences)
for epoch in range(5000):
_, loss = sess.run([optimizer, cost], feed_dict={X:input_batch, Y:target_batch})
if (epoch+1)%1000 == 0:
print(“epoch: “, ‘%04d’%(epoch+1), ‘cost= ‘, ‘%04f’%(loss))

# 预测数据
predict = sess.run([prediction], feed_dict={X: input_batch})
print([sentence.split()[:2] for sentence in sentences], ‘->’, [idx2word[n] for n in predict[0]])

结果打印：

epoch: 1000 cost= 0.008979
epoch: 2000 cost= 0.002754
epoch: 3000 cost= 0.001283
epoch: 4000 cost= 0.000697
epoch: 5000 cost= 0.000406
[[‘i’, ‘love’], [‘i’, ‘like’], [‘we’, ‘love’]] -> [‘damao’, ‘mengjun’, ‘all’]